Bridging the operational AI gap: why integration is the missing link in enterprise AI

Enterprise adoption of AI has progressed considerably. Nearly nine in 10 organizations now use AI in at least one business function, up from roughly half in 2022, and companies with $500 million or more in annual revenue are adopting it more quickly still, according to McKinsey research cited in the report. Executives interviewed for the study distinguish between 2 broad categories of enterprise AI use: everyday applications, which are efficiency focused, such as productivity tools and the automation of repetitive tasks, and transformational projects, which allow companies to take on work that was previously beyond the reach of their workforce.

To understand how organizations are structuring their AI operations, MIT Technology Review Insights surveyed 500 senior IT leaders at US companies with $50 million or more in annual revenue, all of which are pursuing AI in some way.

The results, fielded in December 2025 and supplemented by a series of expert interviews, describe genuine progress alongside a persistent bottleneck.

Three in 4 companies (76%) have at least one AI workflow fully in production. However, 13% report that a piloted project has stalled or been abandoned outright. As the report notes, the real issue is not the AI itself, but the missing operational foundation.

Despite this progress, many companies are still just getting started. Three-quarters of respondents (74%) are still evaluating use cases and tools in at least one department, and 93% are in the piloting stage in one or more departments, indicating that most organizations are working on several fronts simultaneously, with mixed results.

Companies that have a fragmented, siloed application estate with systems that do not connect and legacy systems that cannot scale will struggle to support the additional load from AI.

Ian Thomas, Independent Analyst and Strategy Advisor

Ownership presents a related challenge. Two-thirds of companies (66%) report having no dedicated AI team. Instead, workflow maintenance responsibilities are handled by central IT (21%), departmental operations (25%), or are spread out across the organization with no clear owner (19%). Only about a third of organizations (34%) maintain a team dedicated specifically to AI.

As we move out of the experimentation phase with agentic AI and try to get into those real use cases, one of the top AI investment areas for 2026, according to our research, is data quality and availability.

Shari Lava, Research Vice President, AI and Automation, IDC

Without integration connecting systems, AI tools tend to draw only on the information available within a single application, whether a CRM, an ERP, or a support desk, even though most business processes span multiple systems. This partial view contributes to poor context, unreliable predictions, and AI projects that struggle to progress beyond the pilot stage.

The clearest pattern to emerge from the research is that companies running successful, complex AI projects are, with few exceptions, doing so on top of an integration platform. Ninety percent of companies with an AI workflow fully in production already use one: 37% enterprise-wide, and 53% for specific workflows.

Breadth follows a similar pattern. Four in 10 companies (39%) with an enterprise-wide integration platform are deploying AI across multiple departments at once, compared with 1% or less of companies without one. Integration appears to influence not only the reliability of individual AI projects, but also the extent to which AI initiatives extend across the business rather than remaining confined to a single department.

The problem is most business processes span multiple systems. And to get the full picture, to go through the entire workflow, you need these things to work across the board.

Ronen Vengosh, Chief Strategy Officer, Celigo

The relationship between integration and autonomy is one of the more pronounced findings in the research. Among companies using an enterprise-wide integration platform, 34% describe their AI workflows as mostly autonomous today, compared with 7% of those using integration for specific workflows and 0% of companies with no integration platform at all.

Growing system autonomy is best understood as an iterative process, built on a clean, well-governed data foundation rather than on faith in the model alone. Organizations with enterprise-wide integration also express the greatest confidence about extending autonomy further: nearly a quarter (24%) plan to assign significantly more autonomy to their AI implementations over the next 12 to 18 months, compared with 5% or less among every other group.

Integration platforms allow us to make sure the information is synced between different platforms. They also allow us to identify and bring up alerts of things that are not consistent.

Paula Melo, Vice President of Operational Excellence, Feedzai

Across the interviews conducted for the report, 3 practices distinguish organizations that are moving past the pilot stage from those that remain stalled.

1. Layer AI on top of well-defined, already-automated processes

Nearly half of organizations surveyed (43%) report success applying AI to processes that are already well-defined and automated, a figure that rises to 80% among companies using an enterprise-wide integration platform. Building AI on top of an existing, deterministic automation foundation appears to make new deployments easier to trust and to audit.

2. Avoid an all-or-nothing mindset

AI is not necessarily suited to every use case, or even to every part of a given workflow. Where a process is already well defined and running smoothly on existing automation, applying AI to the entire process tends to be more expensive, slower, and more prone to error.

3. Treat data quality as a prerequisite, not an afterthought

Executives interviewed for the report were consistent on this point: AI cannot simply be layered on top of an existing dataset or system and be expected to perform as intended. Data must be clean, well defined, and well managed before effective integration, and in turn effective AI, can take place. Celigo Ora provides an agentic product surface built on this governed integration foundation.